If a JSON value consists of a large sub-document which is wrapped in
many levels of JSON arrays or objects, serialization of the JSON value
may take a very long time to complete.

This is caused by how the serialization switches between the small
storage format (used by documents that need less than 64KB) and the
large storage format. When it detects that the large storage format
has to be used, it redoes the serialization of the current
sub-document using the large format. But this re-serialization has to
be redone again when the parent of the sub-document is switched from
small format to large format. For deeply nested documents, the inner
parts end up getting re-serializing again and again.

This patch changes how the switch between the formats is done. Instead
of starting with re-serializing the inner parts, it now starts with
the outer parts. If a sub-document exceeds the maximum size for the
small format, we know that the parent document will exceed it and need
to be re-serialized too. Re-serializing an inner document is therefore
a waste of time if we haven't already expanded its parent. By starting
with expanding the outer parts of the JSON document, we avoid the
wasted work and speed up the serialization.